Hacker News Reader

Comment by latchkey | Hacker News Reader

I've run over 120k AMD GPUs. They are ALL snowflakes. Even within the same model number, no two are the same and they can vary quite a bit in performance. It is actually kind of wild to see this at scale. We called it the 'silicon lottery'.

I started to write code to gather data on location on the wafer that it was printed on to see if I could do a correlation there (as I heard a rumor that that the edges aren't as good), but sadly never got a chance to finish it up.

View on HN

The ones from around the edge (you can tell by the distintive curve on one side) are sold as 6900 AF's

The AF (as f**k?) suffix is really powerful: it gives you 1 generation of performance increase.

Depends on the customer I guess. For me I'll always choose AMD for its bulletproof Linux support.

My comment wasn't intended as negative against AMD, just an interesting note about GPU performance measurements.

What does that have to do with their performance variance?

I've noticed this to some extent with their CPUs as well.

From what I can tell this is, in part, because AMD run their chips much closer to their physical limits than say, Intel or Nvidia do. They're all redlining so to speak.

As such this means that any differences in the silicon are much more likely to have an impact on performance.

I can't make sense of this. CPUs microarchitecture is run at their clock speed. If they're running close to their physical limits, either they don't fail in which case all's well, or they fail in which case it's a wipeout such as a blue screen, reboot, freezes, corruption or whatever. AFAICS they can't run correctly but somehow slower. Be good if you could explain where I'm getting it wrong.

> run at their clock speed

The clock speed for a CPU is dynamic. Both Intel and AMD CPUs will boost their clocks to as high as possible given the current conditions.

Some will boost higher than others, even in the same batch. They are even labelled "up to" like a cable internet connection.

You're right, thanks.

Reply on news.ycombinator.com